Re: s390 qemu boot failure in -next

From: Christian Borntraeger
Date: Mon Jun 25 2018 - 04:30:00 EST




On 06/25/2018 10:05 AM, Cornelia Huck wrote:
> On Mon, 25 Jun 2018 09:27:59 +0200
> Christian Borntraeger <borntraeger@xxxxxxxxxx> wrote:
>
>> Also adding QEMU.
>>
>> On 06/25/2018 09:10 AM, Christian Borntraeger wrote:
>>>
>>>
>>> On 06/22/2018 09:47 PM, Guenter Roeck wrote:
>>>> Hi,
>>>>
>>>> starting with commit 's390/boot: make head.S and als.c be part of the
>>>> decompressor only' in -next, s390 immages no longer boot in qemu.
>>>> As far as I can see, the reason is that the command line is no longer
>>>> passed from qemu to the kernel, which results in a panic because the
>>>> root file system can not be mounted.
>>>>
>>>> Was this change made on purpose ? If so, is there a way to get qemu
>>>> back to working ?
>>>
>>> Certainly not on purpose.
>>>
>>> Vasily, I can reproduce this with KVM and an external kernel boot of the vmlinux file (the elf file)
>>>
>>> e.g.
>>>
>>> qemu-system-s390 -enable-kvm -nographic -kernel vmlinux -append "this string no longer is command line"
>>>
>>> The compressed image (bzImage) seems to work fine though.
>>>
>>> This seems to be an unfortunate side effect of QEMUs ways to "guess" its Linux (checking for start
>>> address 0x10000, which is no longer true for the vmlinux file). With the pure vmlinux elf file
>>> the load address is 0x100000 as there is no unpacker.
>
> Do we consider these locations to be an exported interface, or is it
> just QEMU guessing?

It is using what zipl has hardcoded. This works fine for the final image (with the
unpacker), but being able to boot the elf file was just a very nice feature of QEMU.

>
>>>
>>> Guenter, can you check if arch/s390/boot/bzImage works for you as a workaround?
>>
>> Something like this in QEMU
>>
>> diff --git a/hw/s390x/ipl.c b/hw/s390x/ipl.c
>> index f278036fa7..14153ce880 100644
>> --- a/hw/s390x/ipl.c
>> +++ b/hw/s390x/ipl.c
>> @@ -187,11 +187,13 @@ static void s390_ipl_realize(DeviceState *dev, Error **errp)
>> */
>> if (pentry == KERN_IMAGE_START || pentry == 0x800) {
>> ipl->start_addr = KERN_IMAGE_START;
>> - /* Overwrite parameters in the kernel image, which are "rom" */
>> - strcpy(rom_ptr(KERN_PARM_AREA), ipl->cmdline);
>> } else {
>> ipl->start_addr = pentry;
>> }
>> + if (ipl->cmdline) {
>> + /* If there is a command line, put it in the right place */
>> + strcpy(rom_ptr(KERN_PARM_AREA), ipl->cmdline);
>> + }
>
> Check for the magic Linux string (like in the non-elf case) first?

Even that does not exists in vmlinux but only in bzImage with the latest patchset
(in next, but not upstream yet)
>
>>
>> if (ipl->initrd) {
>> ram_addr_t initrd_offset;
>>
>> would put the command line in no matter what the start address is.
>
> I'm for putting that one in (and backporting it to qemu-stable). It's a
> bit worrying, though, that our ipl code is so fragile...

We actually have to combine this with Thomas fix (to check for rom_ptr returning
something sane). It seems that ipl->commandline is always there, so we have to
check for strlen!=0 it seems..

I mean if somebody ask for "-append something" we can certainly always write something
if there is rom/ram.